MiniMax H3: An Omni-Modal Video Model With Native Audio, 2K Output, and Open Weights Coming

MiniMax launched H3 on July 31, 2026, and it is the most interesting video model release in months for one reason: it is not a video model. H3 is a general-purpose omni-modal generation model that takes text, images, video, and audio as input and produces video with native stereo audio at up to 2K resolution and 15 seconds, priced at 0.80 CNY per second at 2K. MiniMax says open weights are coming in the next few days, which would make it the first open-weight model in this tier to ship joint video and audio generation in one forward pass.

Image: MiniMax (via the H3 launch post)

Official Sources#

What Shipped#

H3 accepts a multimodal content array - text, image, video, and audio in any combination - and returns a video clip. The headline capabilities from the launch post:

• Native stereo audio. Voice, sound effects, and music are modeled jointly, not layered on afterward. The output video carries its own soundtrack.
• 2K output by default, up to 15 seconds per clip.
• Accurate text and brand presentation. MiniMax cites instruction following and legible rendered text as first-class outputs, aimed at advertising and e-commerce work.
• V2V motion transfer. A reference video's camera moves and action can be transferred to a new subject.

The generation modes, from the official docs:

Output is MP4 at 2K, 4 to 15 seconds in the MiniMax docs (Vercel lists 5 to 15), in ratios including 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16, or adaptive to a supplied image. Reference and keyframe modes are mutually exclusive.

The Architecture Bet#

Three design choices are worth reading carefully, because they are the difference between "another video upgrade" and a genuine attempt to unify the category:

1. Contextual Omni Representation. MiniMax replaced the "one task, one expert model" split - separate T2I, editing, subject reference, and style reference models - with language as the bridge. The captioning pipeline consumes roughly 100K tokens of source material and distills it to about 4K tokens of structured description. This is how one model handles "reference the camera move in Video 1, have the person in Image 2 sing with the voice from Audio 3" as a single request.

2. H3-VAE. A new tokenizer that delivers a 4x gain in effective sequence length, which is what makes native 2K output affordable without a dedicated upscaler.

3. In-context regeneration. Instead of a super-resolution module, the base model regenerates its own low-resolution output in-context, drawing on the original multimodal context. MiniMax's claim: small text and fine detail recover better than a traditional SR pass can guess.

MiniMax also says it abandoned the Hailuo-02 architecture entirely because it "would introduce unnecessary complexity for a model built around task generalization," and that separating understanding and generation compute lifted training throughput by nearly 30%. This is a deliberate convergence with the FLUX 3 approach: one foundation model, many output modalities.

Pricing, Verified#

From the MiniMax pricing page (fetched July 31, 2026), H3 is billed per second of generated video:

Input material: audio is free, up to 5 images are free (0.20 CNY each beyond that), and input video is billed at the same per-second rate as output at the chosen resolution. For comparison, a 10-second 2K clip costs 8.00 CNY (~$1.11). MiniMax's own framing: at 2K the per-second price is less than a third of mainstream models, and at 768P it is less than half the price of mainstream models at 720P. Treat those relative claims as vendor marketing until independent measurement lands, but the absolute numbers are on the page.

How to Use It#

H3 is an API model, available two ways today.

MiniMax API (async task flow, from the docs):

import os
import requests

api_key = os.environ["MINIMAX_API_KEY"]
headers = {"Authorization": f"Bearer {api_key}"}
BASE_URL = "https://api.minimaxi.com"

payload = {
    "model": "MiniMax-H3",
    "content": [
        {"type": "text", "text": "A white kitten chases a butterfly across a sunlit garden."}
    ],
    "resolution": "2K",
    "duration": 5,
    "ratio": "16:9",
}
resp = requests.post(f"{BASE_URL}/v2/video_generation", json=payload, headers=headers)

The API is asynchronous: create a task, poll for status, download the result from the returned URL.

Vercel AI Gateway, from the changelog:

import { experimental_generateVideo as generateVideo } from "ai";

const { videos } = await generateVideo({
  model: "minimax/minimax-h3",
  prompt: "A white kitten chases a butterfly across a sunlit garden.",
  aspectRatio: "16:9",
  duration: 5,
});

One honest note: this model is not available in OpenCode, and we are not going to fake that section. OpenCode is a coding agent, not a video pipeline, and H3 has no place in it. If you want to wire video generation into an agent workflow, the practical pattern is the one OpenMontage demonstrated: the coding agent owns script, storyboard, and render orchestration, and calls the video API as a tool.

Decision Guide#

H3 is competing with closed video APIs (Veo 3, Sora 2, Kling 2.x) and there are no independent benchmark numbers yet, so the guide is short:

Use H3 when:

• Your output needs synchronized audio. Native stereo in one request removes the separate voice-over and SFX pipeline entirely.
• You want text and brand elements rendered inside the video. That is H3's stated strength and the reason it is aimed at ads and e-commerce.
• Cost is the constraint. 0.80 CNY per second at 2K is a step change from the mainstream closed APIs, if the pricing holds.
• You want a self-hosting escape hatch. Open weights are promised in the coming days; the break-even math changes fast when that happens.

Wait when:

• You need proven reliability at production scale. This is launch day; the closed incumbents have months of hardening.
• Quality parity is unproven. MiniMax's demo clips are impressive but they are demos. Wait for third-party measurement.
• You only need text-to-video with no audio or reference work. Simpler, cheaper models already cover that lane.

FAQ#

What is MiniMax H3?#

An omni-modal generation model launched July 31, 2026. It takes text, images, video, and audio as input and generates video up to 2K resolution and 15 seconds, with native stereo audio, at 0.80 CNY per second.

How much does MiniMax H3 cost?#

0.80 CNY ($0.11) per second of video at 2K, 0.50 CNY ($0.07) at 768P (currently in beta). Up to 5 input images and all input audio are free; extra images cost 0.20 CNY each; input video bills at the same per-second rate.

Is MiniMax H3 open source?#

MiniMax says it plans to open the weights "in the coming days," subject to applicable laws and regulations. No weights or model card are published as of July 31, 2026.

What is the model id?#

MiniMax-H3 on the MiniMax API (POST /v2/video_generation) and minimax/minimax-h3 on Vercel AI Gateway.

How long can generated clips be?#

Up to 15 seconds per clip in MiniMax's docs (4 to 15 seconds on their API, 5 to 15 per Vercel's listing). Generation is asynchronous via task polling.

Sources#

All prices verified July 31, 2026. USD figures are conversions at approximately 7.2 CNY/USD.

Continue Reading#

• FLUX 3: A Unified Multimodal Foundation Model - Black Forest Labs' take on one model for image, video, and audio
• OpenMontage: Agentic Video Production - how coding agents should orchestrate video pipelines
• MiniMax M2.5 Developer Guide - the LLM side of MiniMax's model family
• Kimi K3 Open Weights Release - the open-weights trend H3 joins
• Self-Hosting Open-Weights Models - the economics when weights actually drop

Last updated: July 31, 2026